[improvement] add average input/output token length for hicache benchmark stats output (#10525)

benchmark/hicache/bench_multiturn.py

@@ -479,6 +479,18 @@ class WorkloadGenerator:
             "summary": {
                 "total_requests": len(self.performance_metrics["ttft"]),
                 "request_rate": self.request_rate,
+                "average_prompt_len": (
+                    sum(self.performance_metrics["prompt_len"])
+                    / len(self.performance_metrics["prompt_len"])
+                    if self.performance_metrics["prompt_len"]
+                    else 0.0
+                ),
+                "average_output_len": (
+                    sum(self.performance_metrics["generated_len"])
+                    / len(self.performance_metrics["generated_len"])
+                    if self.performance_metrics["generated_len"]
+                    else 0.0
+                ),
                 "average_ttft": sum(self.performance_metrics["ttft"])
                 / len(self.performance_metrics["ttft"]),
                 "p90_ttft": sorted(self.performance_metrics["ttft"])[
@@ -534,6 +546,12 @@ class WorkloadGenerator:
         print(
             f"  Total requests: {performance_data['summary']['total_requests']} at {performance_data['summary']['request_rate']} requests per second"
         )
+        print(
+            f"  Average Prompt Length: {performance_data['summary']['average_prompt_len']:.2f} tokens"
+        )
+        print(
+            f"  Average Output Length: {performance_data['summary']['average_output_len']:.2f} tokens"
+        )
         print(f"  Average TTFT: {performance_data['summary']['average_ttft']:.2f}")
         print(f"  P90 TTFT: {performance_data['summary']['p90_ttft']:.2f}")
         print(f"  Median TTFT: {performance_data['summary']['median_ttft']:.2f}")